Despite being very loyal to Nvidia in marketing, Nokia has been quietly advancing AI-RAN with its old chip partner Marvell Technology.
Those following AI Wireless Access Network ( AI-RAN ) will have noticed that Nokia has been showcasing some telecom pilot customers in recent weeks, while Nvidia has been hovering in the background like a 'helicopter parent'. After investing $1 billion in Nokia, this major chip manufacturer and its graphics processors ( GPU ) have become at the core of Nokia's RAN strategy. Nokia is designing future 5G and 6G products to run on these GPU , rather than the custom chips currently used in today's networks.
Following these cooperation announcements, one of the three major mobile operators in the UK, Virgin Media O2 (also VMO2), seems to have been merely mentioned in passing. By that time, Nvidia seemed to have quietly withdrawn from the scene like a busy guardian. Strangely enough, VMO2's official statement regarding the AI-RAN experiment did not mention Nvidia, and Nvidia also failed to include VMO2 in the list of telecommunications AI events published after the recent GTC (the GPU Technology Conference GPU Technology Conference) held in Berlin.
Nevertheless, all previous indications suggested that the test site in Paddington, London (Paddington), which uses Nvidia GPU, was running Nokia (AI-RAN) software. However, Nokia later informed Light Reading that this was not the case. Far from using the famous Nvidia chip, this high-capacity site continued to rely on older custom chips to handle new AI-RAN functions. (After new details emerged, Light Reading retracted its initial reports regarding the Paddington tests.)
The statement VMO2 does not mention Nvidia, which could be a clue. Nokia almost never fails to mention this chip manufacturer in statements related to AI-RAN. Around the same time, its competitor British Telecom BT announced its AI-RAN trials with Nokia, and Nvidia appeared in the headlines. Not only did Nokia not mention VMO2 in its AI-RAN summary, but it also remained silent about the custom chips used in AI-RAN and the identity of the supplier.
However, this supplier is actually Marvell Technology. Nvidia is awkwardly squeezing it out of the market through their partnership. Nowadays, Nokia's RAN products, which consume a great deal of resources for the Layer software, rely entirely on Marvell's custom chips. Light Reading has learned that VMO2's AI-RAN experiments specifically used Nokia's internal code-named ABIP Marvell Layer chips.
In the abbreviation “ABI”, “ABI” represents “Airscale Baseband Indoor”, and “P” seems to be just used to distinguish it from an earlier generation of products. “ABIP” may still exist for many years, but according to the current strategy, it will be the last custom chip that Nokia plans to develop. Its successor, “ABIG”, will be based on “Nvidia” technology, where “G” stands for “GPU”.
For Nokia, the potential embarrassment lies in the fact that AI-RAN can be implemented on both existing custom chips and GPU. Since Nokia began collaborating with Nvidia, AI-RAN has almost become synonymous with using Nvidia chips and CUDA software technology. If the existing deployed custom chips also seem to offer the same benefits, it may be more difficult to persuade customers to adopt GPU.
However, it would not be wise to let ABIP lie idle until it becomes obsolete before GPU has been widely deployed. Competitor Ericsson stated that they have already introduced a link adaptation standard, which utilizes the RAN algorithm in their custom chips in approximately 60,000 commercially deployed base stations as a successor solution to AI. The company claims that this measure has increased spectrum efficiency by about 10%. If Nokia is unable to provide competitive software products for the already installed chips, they may lose market share.
This does not necessarily negate the reasons for switching to GPU. Nokia believes that once customers deploy ABIG, they can benefit from years of software upgrades with AI-RAN, continuously improving spectrum efficiency; whereas custom chips require more frequent replacements. After its market share has been declining, and given that Marvell also seems to be concerned about the return on investment in its shrinking RAN industry, Nokia may not have many alternative chip options available.
The problem is that if it turns out that custom chips are still competitive on AI-RAN, operators may be reluctant to spend money on GPU. And the trials of VMO2 seem to indicate this very point.
In one of the busiest areas of the UK, this trial focused on the wireless conditions between antennas and customers' mobile phones in the Paddington area. "The focus of our collaboration with them is to use AI to better estimate channel quality," said Rob Joyce, the Director of Mobile Access Engineering at VMO2, at a recent press briefing.
Traditional algorithms rely on a set of preset parameters regarding these wireless conditions. AI, on the other hand, learns while doing, like an early-blooming apprentice, and is able to quickly adapt to real-time factors. "You will optimize based on what you observe, and it will train itself in its environment," Joyce explains. "Therefore, the base station will adjust its algorithm to that environment."
What VMO2 observed in Paddington was an increase in spectral efficiency by 20%, which means that the site can handle more data without the need for additional capacity or spectrum. The figure of 20% is also consistent with what Nokia recently stated regarding the spectral efficiency improvement they achieved through GPU, which AI-RAN is capable of achieving.
Joyce also acknowledges that Ericsson's AI-RAN products are equally competitive. Ericsson insists on using custom chips, and its CEO, Per Narvinger, has previously stated that using Nvidia's 'native implementation' would lead to a locking effect. 'You don't need GPU to achieve this,' Joyce emphasized at a press briefing. 'To be fair, I think Ericsson's chips are also very good.'
If anyone is not surprised by the implementation of AI-RAN on the Marvell chip, it's Tommi Uitto, the former president of Nokia's mobile network business group. "A few years ago, someone cleverly embedded certain AI and ML [machine learning] capabilities into the chip, but we haven't put them to use yet," he told Light Reading at the beginning of 2025. "Using quite expensive GPU just to improve network performance might not be technically or economically reasonable," he said at that time.
Perhaps the key issue is whether the normal iteration pace of custom chips can keep up with the future software improvements on the GPU platform, whereas the GPU platform does not need to be replaced as frequently as custom chips. Last summer, when discussing that 20% improvement in spectrum efficiency, Nokia believed that it was possible to achieve 50% by next year and double that by 2028.
However, a white paper published by consulting firm Chetan Sharma Consulting and endorsed by Nokia is much more cautious, even raising various concerns. The head of this consulting firm, Chetan Sharma, wrote: "For a site that has already approached the theoretical spectral efficiency limit through traditional methods, there is less room for improvement compared to a site that falls far below the theoretical limit due to interference, channel estimation errors, or poor scheduling."
None of the various laboratory tests and field trials of AI-RAN indicate what baseline was used as a reference for the improvement in spectral efficiency. Experts believe that the RAN algorithm is somewhat akin to a Stradivari violin – after decades of refinement, it is almost impossible to make further improvements. It seems unlikely that AI will be able to create a musical instrument that surpasses this Italian master luthier.
In any case, testing ABIP on AI-RAN makes it seem even less likely that VMO2 will be able to deploy GPU on a large scale. Over the past two years, the company has spent approximately 1.4 billion pounds (1.9 billion US dollars) on upgrading mobile networks, installing new Nokia wireless equipment in southern England and Ericsson equipment in the north. The investment plan of 700 million pounds (940 million US dollars) announced at the beginning of this year does not include the deployment of GPU RAN, as Joyce confirmed to Light Reading.
He also holds a skeptical attitude towards the centralized RAN based on GPU. This approach would concentrate baseband resources in other network facilities, leaving only wireless devices on the towers. “If you have high-capacity fiber optics, you can move GPU to a centralized location,” he said, which would allow for the control of multiple sites. “That’s very advantageous because you can coordinate and schedule between different technologies, and further enhance the effectiveness.”
What Joyce is concerned about are the fibers required for these sites. The Paddington site is connected to 10Gbit/s fiber, which is sufficient to meet local needs. If baseband is to be centralized, VMO2 will have to re-connect this facility to the wireless sites using a technology called CPRI (also known as common public radio interface, a common public wireless interface), or its upgraded version eCPRI. “Each link is 25Gbit/s,” Joyce says. “In the end, you might need to run three such CPRI connections, which means a total of 75Gbit/s to the tower, rather than 10Gbit/s.”
It's almost certain that this means either increasing the fiber optic capacity or paying higher transmission fees, which also explains Joyce's reservations about investing in centralized RAN. "My opinion remains that I believe distributed RAN is the way forward; in this case, if you want to achieve that, you must add additional baseband GPU resources at the site," he said.
For Nokia, the risk lies in customers continuing to resist upgrades that require the deployment of new hardware, regardless of the cost. The AI-RAN features available on Nokia's custom chip products may further diminish customers' motivation to upgrade. Conversely, if today's custom chips are unable to accommodate the emerging AI features already supported by GPU, Ericsson would also face risks. Customers who have benefited from link adaptation improvements on hardware that has been in use for many years may find that they need to replace their devices sooner than expected in order to take advantage of more advanced AI.
The concerns of operators regarding the technical roadmaps of two manufacturers may turn into an opportunity for another to expand its market share. However, due to the very limited number of suppliers in the RAN market, operators are eager for both of these Nordic companies to succeed.
More content:Europe
